Shulou(Shulou.com)11/24 Report--

[guide to Xin Zhiyuan] it is not unreasonable that AI cannot defeat AI. Recently, AI testing tools have failed to give a consistent answer to pictures of infant deaths watched by tens of millions of people.

The AI image detector has been flushed again!

Recently, a large number of photos of conflicts in the Middle East have been released on the public network, showing the world how fragile and helpless life is under such extreme conditions.

Among them, a picture of a "charred baby" is too cruel to believe.

Therefore, someone put the photos into the AI image detector to detect whether the photos were generated by AI.

Sure enough, this photo was identified as "AI generation" by the AI detector, Optic.

On 4chan, there is even an "original picture" in which the original location of the body is actually a dog.

So netizens angrily left a message at the bottom of the publisher's tweet, attacking him for using photos generated by AI to spread false apocalyptic panic.

This tweet believes that the photo was generated by AI and has been read 21 million times in less than 2 days.

But soon netizens found that they put the photos on the same AI detector, and the results were almost random, both AI and humans.

It has been found that as long as the same picture is cropped, or the background color becomes black and white, the detector will assume that the picture was taken by a human.

Sometimes even when the detector flips a coin, the coin will stand up.

So is this picture generated by AI or not?

Finally, AI detector officials also tweeted about this incident, thinking that they had no way to determine whether the picture was generated by AI. I hope everyone will discuss it rationally.

How unreliable is the AI image detector? Professor UC Berkeley Hany Farid, one of the world's leading digital image processing experts, said there was no indication that the image was generated by AI.

"one of the biggest problems with AI image generators is highly structured shapes and lines," Farid said. "if you see the legs and screws and everything looks perfect, then it's almost impossible for the picture to be generated by AI. "

For example, in this famous picture of "SpongeBob made 911", the lines of the twin towers outside the window are not straight, and the dashboards on the plane are twisted into each other.

"We can see in that photo that the structure of the object is accurate, the shadows are accurate, and there are no artifacts-which makes me believe that this picture should be completely real," Fraid said.

Farid also recognized the image through his own other AI image detectors, and the other four AI image detection tools all agreed that the image was not generated by AI.

Farid said, "the AI detector is a tool, but it is only part of the toolkit. Users need to conduct a series of tests on the whole image, and it is impossible to get the answer by pressing a button. "

And the AI detection tool Optic really did not give the details of its own detection technology.

The Optic website also states that "AI detectors may produce inaccurate results."

Professor Farid of AI image detection technology wrote a paper last year, which introduced how to judge the image consistency of AI biograph tool.

By judging the consistency on the image, it can help to determine whether the image is generated by AI.

Links to papers: https://arxiv.org/ abs / 2206.14617?ref=404media.co

The professor first outlines three related physics-based analysis methods, each of which makes use of the same basic perspective geometry inherent in the image formation process.

Vanishing point

Parallel backward lines converge at a vanishing point.

The diagram 1 (a) between the tiles is parallel. When imaging, these lines all converge at a vanishing point. If the parallel line in the scene is deep away from the lens, then there will be a vanishing point, although it may fall outside the image.

If the parallel lines in the scene do not retreat in depth, that is, if they are completely parallel to the lens sensor (at any distance), the parallel lines will be imaged as parallel lines, and for practical purposes, it can be considered that the vanishing point is infinitely far away. This geometry comes from the basic knowledge of perspective projection.

Under perspective projection, the points in the scene (X, Y, Z) are imaged to the points (f X / Z, f Y / Z), where f is the focal length of the lens.

Because the position of the point in the image is inversely proportional to the distance Z, the projection point is compressed as a function of the distance, resulting in the convergence of lines in the image.

two。 The parallel lines on the parallel plane converge to the same vanishing point

The distant box is aligned with the tiles on the floor in figure 1 (b) so that the edge of the box is parallel to the line between the tiles. Because the parallel lines on the parallel plane share a vanishing point, the vanishing point of the side of the box and the tile floor are the same.

3. The vanishing points of all lines on the plane are on the vanishing line.

There are many sets of parallel lines, each converging to a different vanishing point, as shown in figure 1 (c). If parallel line groups span the same plane in the scene, their vanishing point will be on the vanishing line. The direction of the vanishing line is determined by the rotation of the lens relative to the plane crossed by the parallel line.

shadow

Somewhat surprisingly, the same geometry behind Vanishing Point is also suitable for casting shadows.

The image above shows three rays connecting the points on the box and their corresponding points on the cast shadow. After expanding the image boundary, it is found that the three rays intersect at a point that corresponds to the projection of the light source that illuminates the scene.

This geometric constraint related to shadows, objects, and light holds regardless of whether the light source is near (table lamp) or in the distance (sun), and regardless of the position and direction of the surface on which the shadow is cast.

Of course, this analysis assumes that the scene is illuminated by a single main light source, which is evident from the fact that there is only a single cast shadow per object.

In the above example, the light source that illuminates the scene is located in front of the lens, so the projection of the light source is in the upper half of the image plane.

However, if the light is behind the lens, the projection of the light source will be in the lower half of the image plane. Because of this reversal, the shadow of the object constraint must also be reversed.

Therefore, the shadow casting analysis of an image must consider three possibilities:

(1) the light is located in front of the lens, the projection of the light source is located in the upper half of the image plane, and the constraint is anchored on the cast shadow and surrounds the object.

(2) the light is behind the lens, and the light source is projected on the lower half of the image plane, constraining and anchoring the object and surrounding the cast shadow.

(3) the light is located directly above or below the center of the lens, the projection of the light source is at infinity, and the constraints will intersect at infinity. If any of these cases leads to a common intersection of all constraints, casting shadows is physically reasonable.

Reflection

The scene shown in figure 2 below shows three boxes reflected in a flat mirror.

The bottom half of the picture shows the geometric relationship between the real box and the virtual box.

The orange line represents the mirror and is located at the midpoint between the two sets of boxes. The yellow line connects the corresponding points on the real and virtual boxes. The lines are parallel to each other and perpendicular to the mirror.

Now consider how these parallel lines appear when they are superimposed on the scene. Parallel lines are no longer parallel when viewed from the plane of the mirror. Instead, due to perspective projection, these parallel lines converge to a point, just as parallel lines in the world converge to a vanishing point.

Because the lines that connect the corresponding points and their reflections in the scene are always parallel, these lines must have a common point of intersection in the image to be physically reasonable.

Case analysis

Figure 3 above shows three representative examples of AI composite images and analyzes the geometric perspective consistency between the floor and the top of the counter.

Each image (within a few pixels) accurately captures the perspective geometry of the tile floor as evidence of a consistent vanishing point (in blue). However, the vanishing point of the parallel table (presented in cyan) is geometrically inconsistent with the vanishing point of the table.

Align the block accordingly. Even if the Mesa is not parallel to the tile, the cyan vanishing point should be located on the vanishing line (presented in red) defined by the vanishing point of the tile floor. Notice that for the image in the upper-right corner of figure 3, the horizontal lines on the tile floor are almost parallel, so the corresponding vanishing points are at infinity and therefore do not intersect.

Although the vanishing points in these images are locally consistent, they are not globally consistent. The same pattern was found in each of the 25 composite kitchen images.

The image above is a square picture generated with prompts, with obvious inconsistencies in the shadows.

Figure 8 above shows the application of geometric analysis to the image results generated by AI that contain reflections that appear to be fairly accurate.

Although these reflections are visually reasonable, they are not consistent geometrically.

Unlike the casting shadows and geometry in the previous sections, DALL E-2 is difficult to synthesize reasonable reflections, probably because such reflections are not common in their training image data sets.

Based on the understanding of the limitations of AI image generation, through the detection of picture consistency, it can be very helpful to judge whether the picture is composed of AI.

Image recognition is difficult. AI beats AIAI image generator and is constantly evolving.

In the first half of the year, Midjourney went viral and was able to generate enough realistic pictures, but fooled a lot of people.

The 86-year-old pope wore a white melon hat, a white down jacket with a trumpet mouth, a metal cross necklace and a serious expression.

At that time, as soon as the photo was released, it fooled everyone on social media and was forwarded wildly by many netizens, even calling the pope too trendy.

When everyone believed it, someone suddenly pointed out that it was generated by AI, and many people were dumbfounded.

This is just one of the chestnuts, as well as Barra, CEO of Musk's new girlfriend GM, and other fake pictures that have reached the point where they are completely fake.

This incident directly triggered calls for suspension of AI research and development by technology leaders such as Musk and Apple co-founder Stephen Wozniak.

Although AI generation is interesting and convenient, it brings risks to the industry as a whole.

As soon as it is not small, it will be used by people with ulterior motives to spread false information, infringe intellectual property rights, or use it to generate "fruit photos" and so on.

In the next few months, Midjourney will release the latest V6 version, and the current V5 version has done very well in terms of the realism of image generation.

Other AI image generators are also iterating rapidly. Some time ago, OpenAI just released DALL ·E 3, while Microsoft Bing also used DALL ·E 3 for image generation.

Of course, researchers are also trying to build tools that can distinguish images, and the key is how to catch up with the continuous upgrading of AI image generators.

AI testing tools Competition now, more than a dozen companies have provided tools to identify whether images were generated by AI, and their names include Sensity AI (Deep forgery Detection), Fictitious.AI (plagiarism Detection), Originality.AI, and so on.

Artificial intelligence trust and security company Optic has launched a "AI or Not" website.

On this site, you can upload photos or paste image URLs, and the site will automatically determine whether the photos were generated by AI. There is no limit to the number of pictures uploaded.

In addition, you can post or forward a picture on Optic's Twitter account @ optic_xyz, or add # aiornot, and you will get a reply, including the confidence percentage of the image.

Andrey Doronichev, the company's chief executive, says Optic's AI tool can detect artifacts in each image that are invisible to the human eye, such as changes in brightness and color.

Surprisingly, the accuracy of the tool is 95%.

However, with the upgrade iteration of AI image generation tools such as Midjourney, the accuracy of "AI or Not" has dropped to 88.9%.

In this picture of the Pope, for example, AI believes that 87% of the probability is that it was done by humans.

The image of the pope in a white down jacket was fooled before the Optic update. In fact, some netizens said that if you take a closer look at this picture, you will find obvious signs of artificial intelligence generation, including several clearly blurred areas of detail:

A seemingly incomplete hand is trying to grab something that doesn't look like a coffee cup, with a stain next to it.

The cross worn by the pope is not in the shape of a right angle, and it is engraved with a Jesus carved out of clay and sitting.

-the glasses are not consistent with the shadow of the face

All these points indicate that it is generated by artificial intelligence. It only understands the surface of reality, but does not understand the basic rules governing how physical objects interact.

In addition to Optic's tools, Hive, an artificial intelligence company that tags content, recently updated its own free AI-generated content detector.

This AI tool is trained on millions of images of DALL-E, Stable Diffusion, and Midjourney.

Hive predicts that it can accurately detect about 95 per cent of AI-generated images, especially shared images that have gone viral online, often better than other images.

CEO Kevin Guo said that when people share artificial intelligence images, they choose the most realistic fake images, so people can tell what is real.

The image on the left is an image generated by AI, which can be distinguished by two fingers and a strange high-five, while the real image in a normal iStock photo is like the image on the right.

Like Optic, Hive failed to detect the image of Bing Image Creator.

However, these detection tools are not at a standstill, and they will be updated and upgraded as AI images are modeled.

In fact, AI image recognition can not only rely on the detection tools in the industry, but also set up a guardrail during model training.

Many artificial intelligence image generators are also limited to the "blacklist" on which some content can be generated.

For example, Bing Image Creator will mark and block user prompts that require it to create images of well-known public figures.

Midjourney has "human moderators" and is introducing a way to adjust user requests with algorithms.

And according to the DALL E3 technical report, when you ask ChatGPT to generate some "fruit pictures" or pictures involving black and white, the input prompt is directly rewritten.

Add watermarks to AI, big manufacturers are doing in addition, digital watermarking is also one of the important means to enhance the security of generated AI, Microsoft, Google and other technology giants have been used in products.

Microsoft introduced Bing's ability to generate images with the blessing of DALL E 3 at the September Surface conference.

At the same time, to ensure that images are not abused, the Microsoft team uses encryption to generate invisible watermarks for each image, including creation time and date.

Anyone can click on each picture and easily tell if it was generated by AI.

Meta also has open source Stable Signature, which can embed the digital watermark directly into the pictures automatically generated by AI.

Paper address: https://arxiv.org/ pdf / 2303.15435.pdf

It is worth mentioning that the digital watermark generated by Stable Signature is not affected by destructive operations such as cropping, compression, changing color and so on, and can be traced back to the original origin of the image.

It can be applied to diffusion, GAN and other models, such as Stable Diffusion.

And Google has also released a SynthID on Google Cloud Next that watermarks and detects and recognizes images generated by AI.

SynthID uses two deep learning models for watermarking and recognition. They can be trained together on a different set of images.

The combination model is optimized for a series of objectives, including correctly identifying the content with the watermark, and improving the concealment of the watermark by aligning the watermark with the original content intuitively.

The digital watermark generated by SynthID is directly embedded into the pixels of the image, which can not be detected by human eyes. But SynthID can detect and recognize them.

SynthID can help evaluate the possibility that images are created by Imagen. Amit Roy-Chowdhury, a professor of electrical and computer engineering at the University of California, Riverside, said that if we look closely at the background of the image, we can better detect false images with our own eyes.

However, at a time when the iteration of the AI model is accelerating, it is too difficult to have a "golden eye".

Reference:

Https://www.404media.co/ai-images-detectors-are-being-used-to-discredit-the-real-horrors-of-war/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.